Is there a solution for training models to detect CSAM images? I'm imagining some kind of cooperation with the FBI, who I assume have an extensive dataset, where there models could be trained off FBI data without ever sharing the images.

Or, is there some kind one way mask that could be applied to the image such that the FBI could mask their CSAM dataset and release it and people could train detection models on the masked data, then just mask inputs during inference?

I feel like being extremely aggressive about detection and removal is scalable solution here.